Course Matters

1
Course Matters
2
Paper Review
Paper Reviews Due 20 off for each late day
3
Diwali/Deepavali
No Class
4
Guest Lecture
CS6212 Guest Lecture A/V Synchronization SR1
4-6pm
5
Books
6
Project Grading
7
Previously, on CS5248
8
Cello

• OS disk scheduler
• Support multiple classes
• Protect classes from each other
• Work conserving
• Use two-level scheduling

9
CPU vs Disk Scheduling

• Task computation time unpredictable
• Task can be preemptive

10
Problem

• How can multimedia applications co-exists with
  normal applications?

11
Same Old Idea
12
Same Gang

• P. Goyal, X. Guo, and H. M. Vin.
• A hierarchical CPU scheduler for
• multimedia operating systems.
• OSDI96

13
So Far

• Best-Effort Real-Time Scheduling

14
How to Guarantee Services?
15
Resource Reservation CPU
16
Resource Reservations

• How to reserve?
• Overbook?
• Overuse?

17
Example Memory

• How to reserve?
• malloc
• Overbook?
• return NULL
• Overuse?
• crash, throw exception

18
Example Memory

• How to reserve?
• I will need 128KB
• Overbook?
• Sorry!
• Overuse?
• malloc returns NULL

19
Example CPU

• How to reserve?
• Overbook?
• Overuse?

Memory is discrete. CPU time is continuous.
20
How to Reserve?
21
CPU Reservation

• I need C units of time, out of every T units.

22
Two Problems

• How to determine C?
• take measurement and estimate
• What if C is not constant?
• take the maximum

23
Effects of T and C
24
CPU Scheduling Algorithm
25
Classic Real-Time Scheduler
26
Rate Scheduling Algorithms

• Rate Monotonic
• Earliest Deadline First

27
Rate Monotonic
28
EDF
29
Classic Schedulability Theorems

• Rate Monotonic Scheduler
• Earliest Deadline First

30
Assumptions

• programs are periodic
• computation time is constant
• zero context switch time
• programs are independent

31
Why do we care about RM?

• Theoretically, RM has lower CPU utilization (70)
  but
• RM is more efficient than EDF
• Can support up to 80-90 CPU utilization in
  practice
• Dont want to commit 100 CPU anyway (for
  emergency)

32
Rialto Scheduler

• From MS Research

33
Overview

• Given a set of task, pre-compute a scheduling
  graph

34
E
B
C
A
E
B
F
D
E
B
A
E
B
D
35
C
A
E
B
F
D
36
Advantages

• Only need to make new scheduling decisions when
  new task is added

37
Computing Scheduling Graph

• Input A set of tasks with their reservations
• Output A scheduling graph if feasible

38
Many Possible Solutions
C
A
A
E
B
E
B
F
F
D
D
C
39
C
A
F
E
B
F
D
40
Computing Scheduling Graph

• Input A set of tasks with their reservations
• Output A scheduling graph if feasible
• Goals (a) Minimize context switches (b)
  Distribute free time evenly

41
Life is Tough..
NP-Hard
42
Heuristic
10
10
10
10
10
10
10
43
10
5
F
5
5
10
10
10
44
6
1
F
1
BE
1
6
6
6
45
C
4
1
F
1
BE
1
1
1
AD
1
46
E
B
C
E
B
D
A
E
B
F
E
B
A
D
47
Time Constraints
48
Time Constraint

• I need to run the following code, that takes
  30ms, 100ms from now, and the deadline is
  (now180ms).

49
Pseudocode

• can begin_constraint(start_time, deadline,
  estimate)
• if can schedule then
• do work
• else
• adapt
• time_taken end_constraint()

50
Scheduling Constraints
C
A
F
E
B
F
D
51
Scheduling Constraints
C
A
C1, C2, C5
F
E
B
F
D
52
Advantages of Scheduling Graph

• Only need to make new scheduling decisions when
  new task is added
• Feasibility of time constraint can be predicted
  accurately

53
Peaceful Coexistence
54
Running Unreserved Task
C
A
F
E
B
F
D
55
Media Player on Rialto/NT
56
Look inside NT Scheduler

• Scheduling unit Thread
• Priority level 1-15, 16-31
• rt 24 high 13 normal 8 idle 4

57
Priority Boosting

• Example
• Priority 6 after awaken by keyboard input

58
Anti-Starvation

• If hasnt run for 3 seconds
• Boost priority to 14
• Avoid Priority Inversion

59
Mars Pathfinder
60
Rialto/NT
Rialto
Boost Priority to 30
NT Scheduler
61
Win Media Player 6.4
62
Experiment 1

• WMP running alone

63
Experiment 2

• WMP running with priority 8 competitor

64
Experiment 3

• WMP running with priority 10 competitor

65
Experiment 5

• WMP running with priority 10 competitor

66
Experiment 6

• WMP running with priority 10 competitor

67
Experiment 7

• WMP running with priority 10 competitor

68
Summary
69
CPU Scheduling

• With guarantee
• percentage of CPU time
• task with deadline
• Algorithms
• Rate-monotonic
• EDF
• Rialto

70
CS5248 Summary
71
Last Two Lectures
72
Last Two Lectures

• Papers from recent conferences and journals.